OPERADS, ALGEBRAS AND MODULES

J. P. MAY

There are many dierent types of algebra: associative, associative and commutative, Lie, Poisson, etc., etc. Each comes with an appropriate notion of a module. As is becoming more and more important in a variety of elds, it is often necessary to deal with algebras and modules of these sorts up to homotopy. I shall give a very partial overview, concentrating on algebra, but saying a little about the original use of operads in topology. The development of abstract frameworks in which to study such algebras has a long history. As this conference attests, it now seems to be widely accepted that, for many purposes, the most convenient setting is that given by operads and their actions. While the notion was rst written up in a purely topological framework [19], it was thoroughly understood by 1971 [12] that the basic denitions apply equally well in any underlying symmetric monoidal (= tensor) category. The denitions and ideas had many precursors. I will indicate those that I was aware of at the time. Algebraists such as Kaplansky, Herstein, and Jacobson systematically studied algebras dened by dierent kinds of identities. Lawvere [15] formalized algebraic theories as a way of codifying dierent kinds of algebraic structures. Adams and MacLane [16] developed certain chain level concepts, PROPs and PACTs, with a view towards understanding the algebraic structure of the singular chain complex. Stashe [23] introduced A spaces and constructed their classifying spaces, using associahedra and what in retrospect was an example of an operad. Milgram [22] proved an approximation theorem for iterated loop spaces, using what are now known as permutahedra. Boardman and Vogt [3] switched from algebraic to topological PROPs and used PROPs to prove a recognition principle in innite loop space theory. Beck [2] pointed out the relevance of monads to innite loop space theory. Dyer and Lashof [6] systematized homology operations for iterated loop spaces as analogs of Steenrod operations in the cohomology of spaces. It was my extraordinary good fortune to have had close mathematical contact with all of the people whom I have mentioned. I learned from Steenrod at Princeton and from Jacobson at Yale. Remarkably, all of the rest were at Chicago, or were there when the relevant work was done, or were regular visitors there. I wanted a notion that carried the combinatorial structure familiar to me from a paper that I had written [18] that gave a general algebraic approach to Steenrod operations. The diagrams in the denition of an operad action are generalizations of diagrams that were used there to prove the Cartan formula and Adem relations. I wanted the notion to be so intimately related to monads that one could easily go back and forth between locally and globally dened structures. I had a view
1

J. P. MAY

towards using monads to obtain a new approximation theorem for iterated loop spaces and towards using this approximation theorem together with a monadic bar construction to prove a new recognition principle for innite loop spaces. To these ends, I consciously sacriced all-embracing generality: many types of algebras dened by identities are deliberately excluded. The name operad is a word that I coined myself, spending a week thinking about nothing else. Besides having a nice ring to it, the name is meant to bring to mind both operations and monads. Incidentally, I persuaded MacLane to discard the term triple in favor of monad in his book Categories for the working mathematician [17]1, which was being written about the same time. I was convinced that the notion of an operad was an important one, and I wanted the names to mesh. What I did not foresee was just how exible the notion would be, how many essentially dierent mathematical contexts there are in which it would play a natural role, how many philosophically dierent ways it could be exploited. Things have gone so far that I feel quite incompetent to give a thoroughgoing survey. It would be a little like surveying the use of groups in mathematics. Maybe I will feel a little more competent by the end of the conference. 1. The definitions of operads, algebras, and modules To work. Let S be any symmetric monoidal category, with product and unit object . In this talk, I will focus on algebraic examples and their relationship with some basic examples in topology, but there is no point in specializing the general denitions. Details are given in the handout2. While there are perhaps more elegant equivalent ways of writing them down, my original explicit versions of the denitions still seem to be the most convenient, especially for concrete calculational purposes. An operad C consists of objects C (j) with actions by the symmetric groups j , a unit map : k C (1) and product maps : C (k) C (j1 ) C (jk ) C (j) that are suitably associative, unital, and equivariant. The C (j) are thought of as parameter objects for j-ary operations that accept j inputs and produce one output. A C -algebra A is an object together with j -equivariant maps C (j) Aj A that are suitably associative, unital, and equivariant, where Aj denotes the j-fold -power of A with A0 = . An A-module M is an object M together with j1 -equivariant maps C (j) Aj1 M M that are suitably associative, unital, and equivariant. These are the non-negotiable ingredients, but there are variants. If we drop all reference to symmetric groups, we obtain non- versions of the concepts. If
1To quote MacLane (op cit, p. 134), The frequent but unfortunate use of the word triple in this sense has achieved a maximum of needless confusion, what with the conict with ordered triple, plus the use of associated terms such as triple derived functors for functors which are not three times derived from anything in the world. Hence the term monad. 2Reproduced in expanded form earlier in this volume.

OPERADS, ALGEBRAS AND MODULES

we insist that C (0) = , then we call C a unital operad. The resulting map = C (0) A encodes the unit elements of algebras over such operads (and is not to be confused with the unit map : C (1), which encodes the identity operation present on any kind of algebra). Unital operads come with augmentation maps : C (j) and degeneracy maps C (j) C (j 1), 1 i j. Thinking of elements of the C (j) as operations, we think of (c d1 dk ) as the composite of the operation c with the tensor product of the operations ds . When S has an internal Hom functor, so that S (X Y, Z) S (X, Hom(Y, Z)), = this is made precise by the endomorphism operad End(X) of an object X. We set End(X)(j) = Hom(X j , X). The unit is given by the identity map X X, the right actions by symmetric groups are given by their left actions on tensor powers, and the maps are given by composition of tensor products of maps. The identities that dene an operad are then forced by direct calculation. An action of C on A can be redened in adjoint form as a morphism of operads C End(A). The identities that dene a C -algebra are then also forced by direct calculation. There is a notion of a monoid and of a commutative monoid in any symmetric monoidal category S . These are objects A with unit maps A and product maps AA A such that the obvious diagrams commute. There are also notions of left A-modules, which are dened in terms of maps A M M , and of Abimodules. These classical notions are encoded in operad actions. Assume that S has nite coproducts; for a nite set S, let [S] be the coproduct of a copy of for each element of S. Example 1.1. The unital operad M in S has M (j) = [j ]. The unit map is the identity and the product maps are dictated by the equivariance formulas. An M -algebra A is the same thing as a monoid in S . The operad action encodes all of the iterates and permutations of the product on the monoid. In terms of elements, ( a1 aj ) = a(1) a(j) . An A-module M in the operadic sense is an A-bimodule in the classical sense. In terms of elements, given an operad action , we dene am = (e a m) and ma = ( a m), where e and are the identity and transposition in 2 . Conversely, given an Abimodule M , we dene ( a1 aj ) = a(1) a(j) , where j , ai A for 1 i < j and aj M . Example 1.2. The unital operad N has N (j) = for all j. The j -actions are trivial, the unit map is the identity, and the maps are the evident identications. An N -algebra A is the same thing as a commutative monoid in S and an A-module M in the operadic sense is the same thing as a left A-module. We may regard N

J. P. MAY

as a non- operad. In the non- sense, an N -algebra A is a monoid in S , and an A-module is a left A-module in the classical sense. For any unital operad C , the augmentations give a map : C N of operads. Therefore, by pullback along , an N -algebra may be viewed as a C -algebra.

2. Monadic reinterpretation of algebras We recall some standard categorical denitions. Denition 2.1. Let S be any category. A monad in S is a functor C : S S together with natural transformations : CC C and : Id C such that the following diagrams commute:
C / CC o C C CC CC { CC {{ CC {{{ Id C!  }{{ Id C

and

CCC
C

/ CC

 CC

 / C.

A C-algebra is an object A of S together with a map : CA A such that the following diagrams commute: / CA AC CC CC C id C C!  A

and

CCA

/ CA

 CA

 / A.

Taking = , we see that CX is a C-algebra for any X S . It is the free Calgebra generated by X. That is, for C-algebras A, restriction along : X CX gives an adjunction isomorphism C[S ](CX, A) S (X, A), = where C[S ] is the category of C-algebras. Formally, we are viewing C as a functor S C[S ] whose composite with the forgetful functor is our original monad. Thus the monad C is determined by its algebras. Quite generally, every pair L : S T and R : T S of left and right adjoints determines a monad RL on S , but many dierent pairs of adjoint functors can dene the same monad. Now return to our symmetric monoidal category S and assume that it is cocomplete. We have the following construction of the monad of free algebras over an operad C . Denition 2.2. Dene the monad C associated to an operad C by letting CX =
j0

C (j) [j ] X j .

OPERADS, ALGEBRAS AND MODULES

The unit : X CX is id : X = X C (1) X. The product : CCX CX is induced by the following maps, where j = js : C (k) C (j1 ) X j1 C (jk ) X jk  C (k) C (j1 ) C (jk ) X j
id shue

 C (j) X j . Proposition 2.3. A C -algebra structure on an object A determines and is determined by a C-algebra structure on A. Formally, the identity functor on S restricts to give an isomorphism between the categories of C -algebras and of C-algebras. Proof. Maps j : C (j) [j ] Aj A that together specify an action of C on A are the same as a map : CA A that species an action of C on A. Not all monads come from operads. Rather, operads single out a particularly convenient, algebraically manageable, collection of monads. For the operad M , the free algebra M X is the free monoid in S generated by X. For the operad N , the free algebra N X is the free associative and commutative algebra generated by X. Now suppose that C is a unital operad. In this case, there is a monad that is dierent from that dened above but that nevertheless has essentially the same algebras. Since C is unital, a C -algebra A comes with a unit 0 : A. Thinking of as preassigned, it is natural to change ground categories to the category of objects under . Working in this ground category, we obtain a reduced monad C. This monad is so dened that the units of algebras that are built in by the 0 component of operad actions coincide with the preassigned units . Formally, for an object X with unit : X, the reduced monad CX is obtained from CX as an appropriate coequalizer. Informally, elementwise, we identify i (c) y with c si y, where c C (j), y X j , and si = idi1 idji : X j1 X j . Proposition 2.4. Let C be a unital operad. Then a C -algebra structure satisfying = 0 on an object of S under determines and is determined by a C-algebra structure on A. The reduced construction is more general than the unreduced one. Lemma 2.5. For an object X, CX C(X = ) as C -algebras.

In algebraic contexts, there is not much dierence between the two constructions. There, S will be the category of modules over a commutative ground ring k. If X has an augmentation : X k such that = id, then X Ker k = = and therefore CX C(Ker ). In topological contexts, nothing like this works and there is a huge dierence between the two constructions, with the reduced construction being by far the more important one. Modules also admit a monadic reinterpretation: there is a monad C[1] in the category S 2 such that a C[1]-algebra (A, M ) is a C-algebra A together with an

J. P. MAY

A-module M . There is a free A-module functor FA for any C -algebra A, and C[1](X; Y ) is the pair (CX, FCX (Y )). This construction also has a reduced variant. See [14, I4]. 3. The specialization to algebraic operads Let k be a commutative ring and write and Hom for k and Homk . We shall work in the tensor category Mk of Z-graded dierential graded k-modules, with dierential decreasing degrees by one. We implicitly use the standard convention that a sign (1)pq is to be inserted whenever an element of degree p is permuted past an element of degree q. If one prefers the opposite grading convention, one can reindex chain complexes C by setting C n = Cn . I will refer to Z-graded chain complexes over k simply as k-modules. As usual, we consider graded k-modules without dierential to be k-modules with dierential zero, and we view ungraded k-modules as graded k-modules concentrated in degree 0. These conventions allow us to view the theory of generalized algebras as a special case of the theory of dierential graded generalized algebras. When the dierentials on the C (j) are zero, we think of C as purely algebraic, and it then determines an appropriate class of (dierential) algebras. When the dierentials on the C (j) are non-zero, C determines a class of (dierential) algebras up to homotopy, where the homotopies are determined by the homological properties of the C (j). Recall that a map of k-modules is said to be a quasi-isomorphism if it induces an isomorphism of homology groups. Denition 3.1. Let C be a unital operad with each C (j)n = 0 for n < 0. We say that C is acyclic if its augmentations are quasi-isomorphisms. We say that C is -free if C (j) is k[j ]-free for each j. We say that C is an E operad if it is both acyclic and -free; C (j) is then a k[j ]-free resolution of k. By an E algebra, we mean a C -algebra for any E operad C . These were called May algebras when they were introduced by Hinich and Schechtman [10]. If we ignore symmetric groups, we obtain the notion of an A algebra. These are commutative and non-commutative dierential graded algebras up to homotopy. Similarly, there is a class of operads that is related to Lie algebras as E operads are related to commutative algebras, and there is a concomitant notion of a dierential graded Lie algebra up to homotopy, or L algebra. Hinich and Schechtman [11] called these Lie May algebras. An E algebra A has a product for each degree zero element, necessarily a cycle, of C (2). Each such product is unital, associative, and commutative up to all possible coherence homotopies, and all such products are homotopic. There is a long history in topology and category theory that makes precise what these coherence homotopies are. However, since the homotopies are all encoded in the operad action, there is no need to be explicit. In the last lecture of the conference, I will explain a beautiful new way of thinking about A and E algebras which hides the operads in an operadic tensor product. We will there focus on one particular E operad, but all such operads give equivalent classes of A and E algebras. One can treat operads as algebraic systems to which one can apply versions of classical algebraic constructions. An ideal I in an operad C consists of a sequence of sub [j ]-modules I (j) of C (j) such that (c d1 dk ) is in I if either c or any of the ds is in I . There is then a quotient operad C /I with j th k-module

OPERADS, ALGEBRAS AND MODULES

C (j)/I (j). As observed by Ginzburg and Kapranov [9], one can construct the free operad F G generated by any sequence G = {G (j)} of [j ]-modules, and one can then construct an operad that describes a particular type of algebra by quotienting out by the ideal generated by an appropriate sequence R = {R(j)} of dening relations, where R(j) is a sub [j ]-module of (F G )(j). Actually, there are two variants of the construction, one unital and one non-unital. In many familiar examples, called quadratic operads by Ginzburg and Kapranov [9], G (j) = 0 for j = 2 and R(j) = 0 for j = 3. Here, if G (2) is k[2 ] and R(3) = 0, this reconstructs M . If G (2) = k with trivial 2 -action and R(3) = 0, this reconstructs N . In these cases, we use the unital variant. If k is a eld of characteristic other than 2 or 3, we can use the non-unital variant to construct an operad L whose algebras are the Lie algebras over k. To do this, we take G (2) = k, with the transposition in 2 acting as 1, and take R(3) to be the space (F G )(3)3 of invariants, which is one dimensional. Basis elements of G (2) and R(3) correspond to the bracket operation and the Jacobi identity. We shall see shortly that L can be realized homologically by the topological little n-cubes operads for any n > 1. Note that, in these purely algebraic examples, all C (j) are concentrated in degree zero, with zero dierential. The denition of a Lie algebra over a eld k requires the additional relations [x, x] = 0 if char(k) = 2 and [x, [x, x]] = 0 if char(k) = 3. Purely algebraic operads are not well adapted to codify such relations with repeated variables, still less such nonlinear operations as the restriction (or pth power operation) of restricted Lie algebras in characteristic p. The point is simply that the elements of an operad specify operations, and operations by their nature cannot know about special properties (such as repetition) of the variables to which they are applied. As an aside, since in the absence of diagonals it is unclear that there is a workable algebraic analog, we note that a topological theory of E ring spaces has been developed [20, 21]. The sum and product, with the appropriate version of the distributive law, are codied in actions by two suitably interrelated operads. I may say a bit more about this in my second talk. Fix an operad C and a C -algebra A. It is clear that the category of A-modules is abelian. In fact, it is equivalent to the category of modules over the universal enveloping algebra U (A) of A, where U (A) is a certain dierential graded algebra. Denition 3.2. Let A be a C -algebra. The action maps : C (j) Aj1 M M of an A-module M together dene an action map : C(A; k) M = C(A; M ) M. Thus C(A; k) may be viewed as a k-module of operators on A-modules. The free DGA M (C(A; k)) generated by C(A; k) therefore acts iteratively on all A-modules. Dene the universal enveloping algebra U (A) to be the quotient of M (C(A; k)) by the ideal of universal relations. Actually, U (A) can be described more economically as a quotient of C(A; k). Proposition 3.3. The category of A-modules is isomorphic to the category of U (A)-modules. It follows that the free U (A)-module functor gives us the free A-module functor. This implies the following identications of U (A) in special cases.

J. P. MAY

Examples 3.4. (i) For an M -algebra A, U (A) is isomorphic to A Aop . (ii) For an N -algebra A, U (A) is isomorphic to A. (iii) For an L -algebra L, an L-module in our sense is the same as a Lie algebra module in the classical sense, hence U (L) is isomorphic to the classical universal enveloping algebra of L. In my second talk, I shall construct a tensor product on the category of modules over an E algebra A. From the universal enveloping algebra point of view, this should look wholly implausible: a U (A)-module is just a left module, and, since U (A) is far from being commutative, there is no obvious way to dene a tensor product of A-modules, let alone a tensor product that is again an A-module. Moreover, I will give a precise description of U (A) and show that it is quasi-isomorphic to A, as one would hope. Such calculations are often hard to come by, even rationally, as the work of Hinich and Schechtman [11] on L algebras shows. 4. Algebraic operads associated to topological operads An operad E of topological spaces consists of spaces E (j) with right actions of j , a unit element 1 E (1), and maps : E (k) E (j1 ) E (jk ) E (j) such that the appropriate associativity, unity, and equivariance diagrams commute. For deniteness, we assume that E (0) is a point. Via the singular chain complex functor, an operad E of spaces gives rise to an operad C (E ) of (dierential graded) k-modules for any commutative ring k of coecients. The operad E is said to an E operad if each space E (j) is j free and contractible (a universal j -bundle), and C (E ) is then an E operad in the algebraic sense. Similarly, the chain functor C carries E -algebras (= E spaces) to C (E )-algebras and carries modules over an E -algebra to modules over the associated C (E )-algebra. Taking coecients in a eld k, so that we have a Knneth isomorphism, we can u go further and dene homology operads. Denition 4.1. Let E be an operad of spaces. Dene H (E ) to be the unital operad whose jth k-module is the graded k-module H (E (j)), with algebraic structure maps induced by the topological structure maps. For n 0, dene Hn (E ) to be the suboperad of H (E (j)) whose jth k-module is Hn(j1) (E (j)) for j 0; in particular, the 0th k-module is zero unless n = 0. The degrees are so arranged that the denition makes sense. We retain the grading that comes naturally, so that the jth term of Hn (E ) is concentrated in degree n(j 1), but these operads also have evident degree zero translates. If the spaces E (j) are all connected, then H0 (E ) = N and H (X) is a commutative algebra for any E -space X. If the spaces E (j) are all contractible, for example if E is an E operad, then H (E ) = N . Thus, on passage to homology, E operads record only the algebra structure on the homology of E -spaces, although the chain level operad action gives rise to homology operations, including the Dyer-Lashof and Steenrod operations. It is for this reason that topologists did not formally introduce homology operads decades ago. In fact, there is a sharp dichotomy between the calculational behavior of operads in characteristic zero and in positive characteristic. The depth of the original

OPERADS, ALGEBRAS AND MODULES

topological theory lies in positive characteristic, where passage to homology operads jettisons most of the interesting structure. In characteristic zero, in contrast, the homology operads completely determine the homology of the monads E and E associated to an operad E . Here, for a space X, EX = E (j) j X j . For a based space X, EX is the quotient of EX obtained by the appropriate basepoint identications, and EX has a natural ltration with successive quotients E (j)+ j X (j) , where X (j) denotes the j-fold smash power of X. Here the smash product X Y is the quotient of the product X Y obtained by identifying the wedge X Y to a point. The calculational dierence comes from a simple general fact: if a nite group acts on a space X, then, with coecients in a eld k of characteristic zero, H (X/) is naturally isomorphic to H (X)/k[]. This leads to the following result. Theorem 4.2. Let E be an operad of spaces. Let EH denote the monad in the category of k-modules associated to H (E ) and let EH denote the monad in the category of k-modules under k associated to H (E ). If char(k) = 0, then H (EX) EH (H (X)) and H (EX) EH (H (X)) = = as H (E )-algebras for all spaces X, or for all based spaces in the reduced case. This allows us to realize free algebras topologically. Recall that we have the topological (actually, discrete) versions of the operads M and N : M (j) = j and N (j) = {}. For a based space X, M X is the James construction, or free topological monoid, on X, and it is homotopy equivalent to X if X is connected. Similarly, N X is the innite symmetric product, or free commutative topological monoid, on X, and it is homotopy equivalent to the product over n 1 of the Eilenberg-MacLane spaces K(Hn (X), n) if X is connected. Note that the unreduced constructions M X and N X are just disjoint unions of Cartesian powers and symmetric Cartesian powers and are thus much less interesting. At least in characteristic zero, we conclude that H (M X) = MH (H (X)) and H (N X) = NH (H (X)). The functors MH and NH are the free and free commutative algebra functors on unital k-modules that we previously denoted by M and N . Thus the right sides are the free and free commutative algebras generated by H (X). 5. Operads, loop spaces, n-Lie algebras, and n-braid algebras We obtain deeper examples by considering the operads that come from the study of iterated loop spaces. Their homology operads turn out to describe n-Lie algebras and n-braid algebras. Implicitly or explicitly, the case n = 2 has received a good deal of attention in the recent literature of string theory. Although the theorems Im about to state were proven in the early 1970s, their statements came much later, in work of Ginzberg and Kapranov [9] and Getzler and Jones [8]. For each n > 0, there is a little n-cubes operad Cn . It was invented, before the introduction of operads, by Boardman and Vogt [3]. Its jth space Cn (j) consists of j-tuples of little n-cubes embedded with parallel axes and disjoint interiors in

10

J. P. MAY

the standard n-cube. There is an analogous little n-disks operad dened in terms of embeddings of little disks in the unit disk via radial contraction and translation. These are better suited to considerations of group actions and of geometry, but they do not stabilize over n. There is a more sophisticated variant, due to Steiner [24], that enjoys the good properties of both the little n-cubes and the little n-disks operads. Each of these operads comes with a canonical equivalence from its jth space to the conguration space F (Rn , j) of j-tuples of distinct points of Rn . The little n-cubes operad, and any of its variants, acts naturally on all n-fold loop spaces n Y . Since C1 maps by a homotopy equivalence to M , we concentrate on the case n > 1. When k is a eld of characteristic p > 0, the homology of a Cn -space, such as n Y , has an extremely rich and complicated algebraic structure, carrying Browder operations and some of the Dyer-Lashof operations that are present in the homology of E algebras. I will describe the characteristic zero information and a portion of the mod p information in Cohens exhaustive mod p calculations [4, 5]. (There were earlier partial calculations by Arnold [1] in the case n = 2.) Again, we take k to be a eld. Cohens calculations have two starting points. One is his complete and explicit calculation of the integral homology of F (Rn , j), with its action by j , for all n and j. He used this to dene homology operations. The other is my approximation theorem [19]. It asserts that, for a based space X, there is a natural map of Cn spaces Cn X n n X that is an equivalence when X is connected. This allowed Cohen to combine the homology operations with the Serre spectral sequence to compute simultaneously both H (Cn X) and H (n n X) for any X. In characteristic zero, the calculations simplify drastically since calculation of the homology operads H (Cn ) determines H (Cn X). Cohen showed that each n space F (R , j) has the same integral homology as a certain product of wedges of (n 1)-spheres. Therefore, the operad H (Cn ) can be written additively as the reduced sum N Hn1 (Cn ) of its suboperads N and Hn1 (Cn ), where the reduced sum is obtained from the direct sum by identifying the unit elements in N (1) and H0 (Cn (1)). For n 1, the algebras over Hn (Cn+1 ) turn out to be the n-Lie algebras and the algebras over H (Cn+1 ) turn out to be the n-braid algebras. Denition 5.1. An n-Lie algebra is a k-module L together with a map of kmodules [ , ]n : L L L that raises degrees by n and satises the following identities, where deg(x) = q n, deg(y) = r n, and deg(z) = s n. (i) (Anti-symmetry) [x, y]n = (1)qr [y, x]n . (ii) (Jacobi identity) (1)qs [x, [y, z]n ]n + (1)qr [y, [z, x]n ]n + (1)rs [z, [x, y]n ]n = 0. (iii) [x, x]n = 0 if char(k) = 2 and [x, [x, x]n ]n = 0 if char(k) = 3. Of course, a 0-Lie algebra is just a Lie algebra. For a k-module Y and an integer n, dene the n-fold suspension n Y by (n Y )q = Yqn . Proposition 5.2. The category of n-Lie algebras is isomorphic to the category of Lie algebras. There is an operad Ln whose algebras are the n-Lie algebras, and its degree zero translate is isomorphic to L .

OPERADS, ALGEBRAS AND MODULES

11

Proof. For an n-Lie algebra L, n L is a Lie algebra with bracket [n x, n y] = n [x, y]n . Similarly, for a Lie algebra L, n L is an n-Lie algebra. Theorem 5.3. If char(k) = 2 or 3, then, for all n 1, the operad Hn (Cn+1 ) denes n-Lie algebras. That is, Hn (Cn+1 ) Ln . = For a k-module V , let Ln V = n Ln V be the free n-Lie algebra generated by V. Theorem 5.4. Assume that char(k) = 0. For any based space X, (Cn+1 )H (H (X)) H (Cn+1 X) N Ln H (X) = = is the free commutative algebra generated by the free n-Lie algebra generated by H (X). In fact, this can be interpreted as a theorem about n-braid algebras. Denition 5.5. An n-braid algebra is a k-module A that is an n-Lie algebra and a commutative DGA such that the bracket and product satisfy the following identity, where deg(x) = q n and deg(y) = r n. (Poisson formula) [x, yz]n = [x, y]n z + (1)q(rn) y[x, z]n . The Poisson formula asserts that the map dx = [x, ?]n is a graded derivation, in the sense that dx (yz) = dx (y)z + (1)deg(y)deg(dx ) ydx (z). Batalin-Vilkovisky algebras are examples of 1-braid algebras [7], hence the general case, with non-zero dierentials, is relevant to string theory. However, our concern here is with structures that have zero dierential. Theorem 5.6. The homology H (X) is an n-braid algebra for any Cn+1 -space X and any eld of coecients. If char(k) = 2 or 3, then, for all n 1, the algebras over the operad H (Cn+1 ) are exactly the n-braid algebras. Moreover, the free nbraid algebra generated by a k-module V is isomorphic to N Ln V . The last statement could in principle be proven algebraically, but it is much easier to deduce it from the topology, even in characteristic zero. The n-bracket is denoted n and called a Browder operation in the context of iterated loop spaces. The special characteristic 2 and 3 identities (iii) in the denition of an n-Lie algebra are of conceptual interest: they cannot be visible in the operad Hn (Cn+1 ), but they follow directly from the chain level denition of n . 6. Homology operations in characteristic p When C is an E operad, an action of C on A builds in the kinds of higher homotopies for the multiplication of A that are the source, for example, of the Dyer-Lashof operations in the homology of innite loop spaces and the Steenrod operations in the cohomology of general spaces. We describe the form that these operations take in the homology of general E algebras A in this section. Many other examples are known to topologists, such as the Steenrod operations in the

12

J. P. MAY

Ext groups of cocommutative Hopf algebras and in the cohomology of simplicial restricted Lie algebras. We take k = Z and consider algebras A over an integral E operad C . Let Zp = Z/pZ and consider the mod p homology H (A; Zp ). Theorem 6.1. For s 0, there exist natural homology operations Qs : Hq (A; Z2 ) Hq+s (A; Z2 ) and Qs : Hq (A; Zp ) Hq+2s(p1) (A; Zp ) if p > 2. These operations satisfy the following properties (1) Qs (x) = 0 if p = 2 and s < q or if p > 2 and 2s < q. (2) Qs (x) = xp if p = 2 and s = q or if p > 2 and 2s = q. (3) Qs (1) = 0 if s > 0, where 1 H0 (A; Zp ) is the identity element. (4) (Cartan formula) Qs (xy) = Qt (x)Qst (y). (5) (Adem relations) If p 2 and t > ps, then Qt Qs =
i

(1)t+i (pi t, t (p 1)s i)Qs+ti1 Qi ;

if p > 2, t ps, and denotes the mod p Bockstein, then Qt Qs =

(1)t+i (pi t, t (p 1)s i)Qs+ti Qi

(1)t+i (pi t 1, t (p 1)s i)Qs+ti Qi ;

(i + j)! if i 0 and j 0 (where 0! = 1), and (i, j) = 0 if i i!j! or j is negative; the sums run over i 0. here (i, j) = For the proof, one simply checks that one is in the general algebraic framework of a 1970 paper of mine [18] that does the relevant homological algebra once and for all. That paper should have been about operad actions. As I mentioned, however, it was actually written shortly before I invented operads. The point is that C (p) is a p -free resolution of Z, so that the homology of C (p) p Ap is readily computed, and computation of : H (C (p) p Ap ; Zp ) H (A; Zp ) allows one to read o the operations. The Cartan formula and the Adem relations are derived from special cases of the diagrams in the denition of an operad action via calculations in the homology of groups. Notice the grading. The rst non-zero operation is the pth power, and there can be innitely many non-zero operations on a given element. This is in marked contrast with Steenrod operations in the cohomology of spaces, where the last non-zero operation is the pth power. In fact, Steenrod operations are dened on cohomologically graded E algebras that are concentrated in positive degrees, in which context the complexes C (j) of the relevant E operad must be regraded as cochain complexes concentrated in negative degrees. If we systematically regrade homologically, then Dyer-Lashof and Steenrod operations both t into the general context of the theorem, except that the adjective Dyer-Lashof is to be used when the underlying chain complexes are positively graded and the adjective Steenrod is to be used when the underlying chain complexes are negatively graded.

OPERADS, ALGEBRAS AND MODULES

13

7. A conversion theorem In characteristic zero, E operads carry no more homological information than the operad N , and similarly for more general types of algebras. Lemma 7.1. Let : C P be a quasi-isomorphism of operads over a eld k of characteristic zero, such as the augmentation : C N of an acyclic operad. Then the map CX P X induced by is a quasi-isomorphism for all k-modules X. Taking P = N and P = L , this leads to a proof that, when k is a eld of characteristic zero, E algebras are quasi-isomorphic as E algebras to commutative DGAs and L algebras are quasi-isomorphic as L algebras to dierential graded Lie algebras, and similarly for modules. That is, one can convert operadic algebras and modules to genuine algebras and modules without loss of information. The proof is a typical exercise in the use of the two-sided monadic bar construction that I introduced in the same paper [19] in which I rst dened operads. Theorem 7.2. Let k be a eld of characteristic zero and let : C P be a quasiisomorphism of operads of k-complexes. Then there is a functor W that assigns a quasi-isomorphic P-algebra W A to a C -algebra A. There is also a functor W that assigns a quasi-isomorphic W A-module W M to an A-module M . An acyclic operad augments by a quasi-isomorphism to the operad N that denes commutative DGAs, so the following consequence is immediate. Corollary 7.3. Let k be a eld of characteristic zero and let C be an acyclic operad of k-complexes. Then there is a functor W that assigns a quasi-isomorphic commutative DGA W A to a C -algebra A. There is also a functor W that assigns a quasi-isomorphic W A-module W M to an A-module M . The following consequence for dierential graded Lie algebras is also immediate. Corollary 7.4. Let k be a eld of characteristic zero and let J be an operad with a quasi-isomorphism to L . Then there is a functor W that assigns a quasiisomorphic dierential graded Lie algebra W L to a J -algebra L. There is also a functor W that assigns a quasi-isomorphic W L-module W M to an L-module M . The second corollary may be of interest in applications to string theory. The motivation for the rst corollary originally came from algebraic geometry, where it answered a question of Deligne concerning mixed Tate motives. The techniques of proof also apply to replace partial algebras over operads by genuine algebras over operads, in any characteristic. This has real force. Using it, Kriz and I were able to carry out a suggestion of Deligne for the construction of categories of both integral and rational mixed Tate motives [13, 14]. However, that is a subject for another talk. I hope that this has given a bit of a feel for some of the ways that operads work, and I look forward to learning more from all of you during the conference. References
[1] V. I. Arnold. Cohomology of the group of dyed braids. Math. Zametki 5(1969), 227-231. [2] J. Beck. On H-spaces and innite loop spaces, in Springer Lecture Notes in Mathematics Vol. 99. Springer, 1969.

14

J. P. MAY

[3] J. M. Boardman and R. M. Vogt. Homotopy invariant structures on topological spaces. Springer Lecture Notes in Mathematics Vol. 347. Springer, 1973. [4] F. R. Cohen. PhD thesis, University of Chicago, 1972. [5] F. R. Cohen, T.J. Lada, and J. P. May. The Homology of Iterated Loop Spaces. Springer Lecture Notes in Mathematics Vol. 533. Springer, 1976. [6] E. Dyer and R. K. Lashof. Homology of iterated loop spaces. American J. Math. 84(1962), 35-88. [7] E. Getzler. Batalin-Vilkovisky algebras and two-dimensional topological eld theories. Comm. Math. Phys. 159(1992), 265-285. [8] E. Getzler and J. D. S. Jones. n-Algebras and Batalin-Vilkovisky algebras. Preprint, 1992. [9] V. A. Ginzburg and M. M. Kapranov. Koszul duality for operads. Duke Math. J. 76(1994), 203272. [10] V. A. Hinich and V. V. Schechtman. On homotopy limit of homotopy algebras, in Springer Lecture Notes in Mathematics Vol 1289. Springer, 1987. [11] V. A. Hinich and V. V. Schechtman. Homotopy Lie algebras. Advances in Soviet Mathematics, 16(1993), 127. [12] G. M. Kelly. On the operads of J. P. May. Preprint, University of Chicago, January 1972. [13] I. Kriz and J. P. May. Derived categories and motives. Mathematical Research Letters, 1(1994), 8794. [14] I. Kriz and J. P. May. Operads, Algebras, Modules, and Motives. Astrisque Vol 233. 1995. e [15] F. W. Lawvere. Functional semantics of algebraic theories. Proc. Nat. Acad. Sci. USA 50(1963), 869-872. [16] S. MacLane. Categorical algebra. Bull. Amer. Math. Soc. 71(1965), 40-106. [17] S. MacLane. Categories for the Working Mathematician. Springer-Verlag, 1976. [18] J. P. May. A general approach to Steenrod operations, in Springer Lecture Notes in Mathematics Vol 168. Springer, 1970. [19] J. P. May. The Geometry of Iterated Loop Spaces. Springer Lecture Notes in Mathematics Vol. 271. Springer, 1972. [20] J. P. May. E ring spaces and E ring spectra. Springer Lecture Notes in Mathematics Vol. 577. Springer, 1977. [21] J. P. May. Multiplicative ininite loop space theory. J. Pure and Applied Algebra, 26(1982), 169. [22] R. J. Milgram Iterated loop spaces. Annals of Mathematics 84(1966), 386-403. [23] J. D. Stashe. Homotopy associativity of H-spaces I, II. Trans. Amer. Math. Soc. 108(1968), 457-470. [24] R. Steiner. A canonical operad pair. Math. Proc. Camb. Phil. Soc., 86(1979), 443449. The University of Chicago, Chicago, IL 60637 E-mail address: may@@math.uchicago.edu